We need a two-level sparse tree for best efficiency in this case; the first split uses TypeEnvironment as a key, under which we store an additional Map to further organize by K. This used to be implemented as a HashMap with a composite key {TypeEnvironment,K}, but this wasn't memory efficient as all composite keys take too much space: the nested structure has some overhead as well, but has been shown to cost less overall. In addition, we strife to not need to compute hashkeys multiple times while not needing to allocate stateful lambdas; for this reason the tree nodes need to store some additional references.